nasa-ibm-ai4science/Sombench-pretraining-data

SomBench Pre-training Corpus: Multimodal Lunar Tiles

Dataset Summary

This includes a small sample from SomBench: a corpus of co-registered, multimodal lunar image tiles built for large-scale self-supervised (foundation-model) pre-training. It contains a subset of modalities from the low-resolution (WAC-anchored) and high-resolution (NAC-anchored) tracks specifically used in pretraining.

Tiles are anchored to individual LROC Experiment Data Record (EDR) image canvases rather than to a fixed map grid. A sliding window enumerates fixed-shape 512 × 512-pixel patches over each EDR, and every co-registered modality is snapped to that same per-tile boundary. Each tile therefore bundles many physically distinct measurements (optical, topographic, spectral, radar, thermal, gravity) over the same ground patch, covering the same spatial bounds. Anchoring to image canvases also preserves the optical observation context, and overlapping EDR coverage exposes models to the same terrain under different illumination conditions.

The corpus is organized into two parallel tracks that share construction logic but differ in their optical anchor and ground scale:

• —`WAC_LowRes/`: anchored to LROC Wide-Angle Camera (WAC) visible imagery at 100 m/pixel; each tile covers ≈ 51.2 × 51.2 km.
• —`NAC_HighRes/`: anchored to LROC Narrow-Angle Camera (NAC) imagery at 1 m/pixel; each tile covers ≈ 512 × 512 m.

Both tracks use a 512 × 512-pixel tile size for the anchor modality (WAC or NAC). Coarser modalities are resampled, while preserving native resolution as much as possible, to align to the exact bounds of the anchor modality.

Note: The full dataset is available on AWS at s3://nasa-lunar-fm-bench/ and can be accessed using the AWS CLI:

aws s3 ls s3://nasa-lunar-fm-bench/ --no-sign-request

Supported Tasks and Applications

• —Self-supervised / foundation-model pre-training: masked image modeling and multimodal SSL over co-registered layers.
• —Multimodal representation learning and cross-modal fusion: joint embeddings across optical, topographic, spectral, radar, thermal, and gravity modalities.
• —Multi-scale learning: pairing meter-scale NAC context with hundreds-of-meters-to-kilometer static layers within one framework.
• —Downstream fine-tuning: the test split is reserved for evaluation and fine-tuning on the SomBench application benchmarks (crater detection, IMP segmentation, ice prospectivity).

Tracks

Directory Layout

Sombench-pretraining-data/
├── WAC_LowRes/            # low-resolution (WAC-anchored)
    ├── WAC_LowRes.parquet
    ├── aspect/
    ├── dtm/
    ├── metadata/
    ├── slope/
    ├── uv/
    └── vis/
├── NAC_HighRes/           # high-resolution (NAC-anchored)
    ├── NAC_HighRes.parquet
    ├── aspect_3m/
    ├── dtm_3m/
    ├── metadata/
    ├── nac/
    └── slope_3m/

The .parquet file in each track has a row for each tile and includes the paths to the paired modalities (e.g. DTM_TILE), metadata associated with the EDR image (e.g. INCIDENCE_ANGLE), and DATASET assignment for training.

Within each modality sub-directory, tiles are stored as one netCDF file per patch, named by the source EDR product id and the sliding-window row/column:

{product_id}_r{row}_c{col}.nc

Modalities

Each modality lives in its own sub-directory (the lowercase "dataset key"). The table lists source, native resolution, and band content. Availability differs by track and by region: some layers exist only at the poles, others only outside the poles. Band counts are structural (from the SomBench paper); native resolutions are the upstream product resolutions before per-tile snapping.

NOTE: This table includes all modalities that are included in SomBench; however, the HuggingFace sample inlcudes only modalities used in pre-training, including 6 WACLowRes modalities (`vis`, `uv`, `dtm`, `aspect`, `slope`, `metadata`) and 5 NACHighRes modalities (nac, dtm, aspect, slope, metadata).

Regional availability. Most modalities are present in both polar and non-polar tiles wherever data exists. The exceptions: sp_mineralogy, tbol_poles, dice, psr, avg_illum, albedo, and hydrogen are polar-only, while tio2, hpar, and rock_abundance are non-polar-only. A sub-directory may still appear in both tracks even where its tiles are populated for one region only.

Tile File Format

Every tile is a self-describing netCDF4 file (written via h5netcdf with Bitshuffle + LZ4 chunk compression; one chunk per variable):

• —Coordinate vectors: 1-D x and y pixel-center coordinates in the tile's CRS.
• —Global attributes: full CRS as a WKT string (crs), per-axis pixel resolution (pix_res_x, pix_res_y), and a comma-separated list of band names (band_names).
• —Value-range clipping applied at write time from a per-modality registry (e.g. slope ∈ [0°, 90°], normalized reflectance ∈ [0, 1], albedo ∈ [0, 1], TiO₂ ∈ [0, 100] %). NaN encodes a genuine data gap and is preserved through clipping; the categorical geomap layer is exempt so class indices are unchanged.
• —Snapped resolution: because coarse modalities are resampled to align with the anchor grid, the stored pixel size can differ slightly from native (e.g. 60 m → ≈ 56.9 m in the NAC track; 100 m → ≈ 102.4 m). The per-tile CRS and resolution are always recorded in the file's attributes.

Multi-band / multi-phase notes

• —tbol: all 24 sub-solar-longitude snapshots (15° steps) plus a duplicated "closest" band → 25 bands.
• —tbol_poles: 24 phases × summer/winter (48) plus a "closest" band → 49 bands (polar tiles only).
• —Multi-band spectral/mineral layers (mi_norm_ref, mi_mineralogy, wac_norm_ref, sp_mineralogy, sw_fe) are stacked along a band dimension; if any band is missing, the whole layer is recorded as missing for that tile.

Catalogs and Metadata

Each track has a Parquet catalog (`WAC_LowRes.parquet` and `NAC_HighRes.parquet`) with one row per tile and these columns:

• —Identification: PRODUCT_ID (source EDR id), ROW, COL (sliding-window offset), LTM_CODE (LTM zone or LPS_N/LPS_S).
• —Per-modality: {MODALITY}_TILE, relative path to the tile's netCDF file (e.g. WAC_VIS_TILE, NAC_TILE, DTM_60M_TILE, TBOL_TILE, …) and {MODALITY}_FRACTION_NULL, the tile's NaN fraction for that layer. (METADATA_TILE, and low-res CRATERS_TILE, have no null-fraction column.)
• —Geometry: CENTER_LATITUDE/CENTER_LONGITUDE, corner UPPER_LEFT_* / LOWER_RIGHT_* lon/lat, and projected bounds BOUNDS_XMIN/XMAX/YMIN/YMAX.
• —Viewing geometry: INCIDENCE_ANGLE, EMISSION_ANGLE, PHASE_ANGLE, SUB_SOLAR_GROUND_AZIMUTH, SUB_SOLAR_LATITUDE, SUB_SOLAR_LONGITUDE.
• —Split & flags: DATASET (train/val/test) and ALLOW_NANS_OPTICAL_DTM_SLOPE_ASPECT, a boolean marking tiles permitted to contain NaN in the optical / DTM / slope / aspect layers (primarily polar tiles admitted with partial data; 4,588 in the low-res track, 3,888 in the high-res track).

Loading on the Hub. load_dataset(...) and the dataset viewer return this catalog (per-tile paths and metadata), not the netCDF imagery. Fetch each tile from the path in its {MODALITY}_TILE column (e.g. via huggingface_hub).

The METADATA_TILE column (and the metadata modality directory) carries associated per-tile metadata (illumination geometry, location, etc.).

Splits

Splits are assigned at the grid-cell (LTM zone) level to prevent spatial leakage:

• —Zone partition: unique LTM_CODE grid cells are randomly partitioned into train / validation / test at a target of 75 % / 15 % / 10 % of cells. Whole zones go to a single split, so nearby tiles never straddle splits.
• —Buffer exclusion: tiles that straddle two zones are dropped to guarantee strict spatial separation.
• —Zone/partition assignments are applied consistently across both tracks.

Released split sizes (DATASET column of the catalogs):

Train + validation are intended for pre-training; test is reserved for downstream evaluation and fine-tuning. Because zones vary in tile count, the realized tile ratio differs from the 75/15/10 cell ratio.

Coverage and Projections

Tiles span from the equator to ±82° latitude in 90 Lunar Transverse Mercator (LTM) zones plus two Lunar Polar Stereographic (LPS) caps (|φ| ≥ 82°), using WKT definitions from McClernan et al. (2025). Non-polar windows are rejected if the required anchor modality contains any NaN; polar windows are admitted with any valid data to maximize coverage, so polar tiles more often contain partial NaN. Tiles that straddle two zones are dropped, keeping split boundaries clean.

Known Limitations

Heterogeneous spatial resolution, illumination-driven appearance changes, label scarcity/class imbalance, and tens-of-meters absolute geolocation uncertainty all apply. In practice: expect resampling artifacts where coarse modalities are snapped to the anchor bounds; treat NaN as a genuine data gap (especially in polar tiles, which are admitted with partial coverage); and use the ALLOW_NANS_OPTICAL_DTM_SLOPE_ASPECT flag when building NaN-free training sets. See the SomBench paper for the full discussion and recommended practices.

Citations

@article{fraccaro2026lfm,
  title  = {Multimodal-Multiresolution Foundation Model for Lunar Remote Sensing},
  author = {Fraccaro, Paolo and Nyirjesy, Gabby and Szwarcman, Daniela and Patil, Himanshu
            and Gaur, Vishal and Lal, Rohit and Slank, Rachel A. and Dawson, Geoffrey
            and Debary, Hiyam and Dionelis, Nikolaos and Barker, Michael K. and Annex, Andrew
            and Viswanathan, Vishnu and Morse, Zachary and Schaefer, Ethan I. and Kumar, Ankur
            and Watson, Campbell D. and Dawson-Rigas, Rebekah I. and Maskey, Manil
            and Roy, Sujit and Ramachandran, Rahul and Bernab\'e-Moreno, Juan},
  year   = {2026}
  howpublished = {\url{https://huggingface.co/collections/nasa-ibm-ai4science/nasa-ibm-lunar-fm-and-downstream-models}}
}

@misc{sombench2026collection,
  author = {Patil, Himanshu and Nyirjesy, Gabby and Slank, Rachel A. and Gaur, Vishal
          and Szwarcman, Daniela and Fraccaro, Paolo and Dionelis, Nikolaos and Barker, Michael K.
          and Annex, Andrew and Viswanathan, Vishnu and Morse, Zachary and Schaefer, Ethan I.
          and Debary, Hiyam and Kumar, Ankur and Lal, Rohit and Dawson, Geoffrey
          and Watson, Campbell and Dawson-Rigas, Rebekah I. and Maskey, Manil
          and Bernab\'e-Moreno, Juan and Ramachandran, Rahul and Roy, Sujit},
  title        = {{SomBench}: Benchmark Dataset for Advancing Machine Learning in Lunar Science},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/collections/nasa-ibm-ai4science/lunar-fm-ml-ready-benchmark-dataset-sombench}}
}

License

Released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.